Author: Ming Dong, Carri K. Glide-Hurst, Joshua Pan, Nicholas R. Summerfield 👨🔬
Affiliation: Department of Computer Science, Wayne State University, Departments of Human Oncology and Medical Physics, University of Wisconsin-Madison, Department of Human Oncology, University of Wisconsin-Madison 🌍
Purpose: Radiation dose to the cardiac nodes is more strongly associated with conduction disorders and arrythmias than whole heart (WH) metrics. However, node segmentation is challenging due to complex anatomical variations and poor visualization. Current solutions implement atlas and geometric models. We tackle this complex problem using a deep learning (DL) approach, leveraging soft-tissue contrast from MR-Linac data for automatic cardiac node segmentation incorporating cardiac substructures (CS) to inform predictions, yielding patient-specific node models.
Methods: 18 patients underwent radiotherapy on a 0.35T MR-linac and were retrospectively evaluated. Manually identified sinoatrial and atrioventricular nodes were used to train (n=10), validate (n=3), and test (n=5) a novel 2-channel (Images+Delineations) nnU-Net. Thirteen identified CS (WH/Chambers/Great Vessels/Coronary Arteries) were used as complimentary input information. A sensitivity test was designed including MR Image-only, Image+WH, Image+6CS (Chambers/Superior Vena Cava/Ascending Aorta), and Image+12CS (Chambers/Great Vessels/Coronary Arteries). A trained/validated nnU-Net with dual self-distillation was leveraged during testing/validation, predicting CS and conserving autonomy. Finally, geometric derivation from predicted CS was performed following literature as a baseline comparison. Performance was compared using the Dice Similarity Coefficient (DSC), 95% Hausdorff distance (HD95), and qualitative assessment.
Results: No differences were observed between the Image+6CS (DSC/HD95: 0.53±0.18/7.1±2.6mm) and Image+12CS (0.53±0.19/7.1±2.8mm) models. Incorporating CS into the model improved predictions against Image+WH (0.44±0.26/16.4±28.4mm), Image-only (0.24±0.21/16.5±11.3mm), and geometric (0.43±0.16/8.6±2.9mm) methods. Image+6CS scored 0.69±0.09/4.7±1.1mm and 0.38±0.03/9.4±0.9mm respectively for the sinoatrial and atrioventricular nodes. Using 6CS provided equivalent performance to 12CS while image+WH and image-only models lacked enough information to form strong predictions. Qualitative assessment revealed the Image+6 model predicting the most localized nodes.
Conclusion: DL with complimentary structural CS input successfully localized and identified the cardiac nodes. With further validation and translation to other datasets, integrating cardiac node segmentation into radiotherapy treatment planning will enable more advanced cardiac sparing and support future studies evaluating radiation-induced conduction disorders.